Physics‑Informed Learning Meets Classical Model-Based Control for High‑DoF Robot Manipulators

About the Project

Robotic manipulators with a high number of degrees of freedom epitomize systems whose rich nonlinear dynamics and strict performance requirements challenge both classical control theory and data‑driven methods. This PhD project seeks to traverse the bridge between these paradigms by developing, integrating, and rigorously comparing three families of controllers—model‑reference adaptive laws, sliding‑mode robust schemes, and Model Predictive Control (MPC) enhanced with Physics‑Informed Neural Networks (PINNs). Experiments will be carried out in our Robotics Lab on a 7-DoF Franka Research 3 (a.k.a. Franka Emika Panda) robotic manipulator or a 6-DoF industrial Fanuc M-10iA.

Building on the foundational work of Nicosia and Tomei in model‑reference adaptive control for manipulators (MRAC) [1], the candidate will re‑implement classic adaptive schemes, demonstrating guaranteed convergence and robust tracking despite substantial parameter uncertainty. Sliding‑mode controllers—renowned for their disturbance rejection—will complement this baseline, highlighting practical issues such as control chattering and adaptation speed [2].

In parallel, the project will explore MPC augmented by PINN‑based surrogates [3]. The student will train a neural network that embeds the manipulator’s differential equations into its architecture, then use it as the predictive model inside an MPC loop. Leveraging automatic differentiation, this approach aims to vastly accelerate gradient computations compared to numerical integration, as demonstrated in soft‑robotics contexts [4]. The result is an MPC formulation capable of enforcing constraints and planning over horizons at near‑real‑time rates. The controllers will be implemented on real industrial manipulators: the experiments will introduce systematic perturbations—link masses, payloads, joint friction—to compare each method’s tracking accuracy, robustness margins, computational overhead, and tuning complexity. This comparative study will yield clear insights into where classical methods excel, where learning‑informed surrogates add value, and how they might be synergistically combined.

Although adaptive MPC and learning‑informed controllers have each been explored independently, their fusion remains underexplored for high‑DoF arms. Tube‑based sliding‑mode MPC frameworks [5] hint at combining robustness and constraint handling; embedding PINN surrogates into such architectures could further enhance performance under severe model mismatch.

This doctoral research combines the proven reliability of adaptive and sliding‑mode control with the practical flexibility and speed of physics‑informed learning. Candidates will gain expertise across MATLAB, Python, deep‑learning frameworks (TensorFlow/PyTorch), automatic differentiation, and hands‑on robotics. Deliverables—open‑source implementations, benchmark studies, and novel hybrid control designs—will directly inform both academic inquiry and industrial application in manipulators and autonomous systems.

Embedding advanced control methods into real‑world robotics carries clear societal and economic benefits. By uniting adaptive laws, MPC and physics‑informed neural surrogates, this project promises manipulators that track complex trajectories more accurately, require less calibration, and accommodate payload or model changes. In an industrial setting, these improvements translate directly into reduced downtime, lower maintenance costs, and faster production cycles—benefits that cascade through supply chains to lower the price of manufactured goods.

Moreover, enhanced controller robustness means robots can safely work alongside humans in manufacturing, logistics or even healthcare, reducing workplace injuries and enabling more flexible “cobotic” deployments in small and medium‑sized enterprises. Energy‑efficient trajectories and smoother actuation—by avoiding aggressive torque spikes or conservative safety margins—also reduce power consumption, cutting operational expenses and helping factories meet increasingly stringent environmental targets.

Overall, this project offers clear real‑world benefits—reducing costs, saving energy, and improving workplace safety—while showcasing cutting‑edge advances in robotics, control, and machine learning, making it an appealing opportunity for candidates passionate about both technological excellence and practical impact.

Funding Notes

there is no funding for this project

References

[1] S. Nicosia, P. Tomei, “Model reference adaptive control algorithms for industrial robots,” Automatica, 20(5), 1984.
[2] V. Utkin, “Sliding Modes in Control and Optimization”. Springer-Verlag, 1992.
[3] Raissi, M, Perdikaris, P., Karniadakis, G.E.,” Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations”, Journal of Computational Physics, 2019.
[4] M. Lahariya, C. Innes, C. Develder, and S. Ramamoorthy, “Learning physics-informed simulation models for soft robotic manipulation: a case study with dielectric elastomer actuators,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Kyoto, Japan, 2022.
[5] Zhang J, Li H, Liu Q, Li S. “The model reference adaptive impedance control for underwater manipulator compliant operation,” Transactions of the Institute of Measurement and Control, 2023.